IDDF2022-ABS-0265 Deep learning for predicting postoperative recurrence of colorectal cancer based on histological images

BackgroundAccurate estimation of the risk of postoperative recurrence in patients with non-metastatic colorectal cancer (CRC) is crucial to improve the long-term prognosis. However, the accuracy of the existing models for predicting postoperative recurrence is limited. Deep learning is expected to automatically extract a large number of microscopic features of CRC from the whole slide imaging (WSI), so as to provide more abundant prognostic information. Therefore, we sought to develop a deep learning model based on histological images to stratify patients with CRC with different recurrence risks.MethodsWe collected 729 (2919 WSIs) cases with non-metastatic CRC who had undergone radical surgery from three hospitals, including 553 (2215 WSIs), 76 (328 WSIs) and 100 (376 WSIs) cases for model development and as first and second external test sets respectively to evaluate model performance. Patients who relapsed within 5 years after the operation was classified as recurrence group, while others were assigned to non-recurrence group. The deep learning model was trained by labeled images and could output the recurrence risk score (DL-RRS) by the feature it has learned on the training set. AUC were measured for model performance.ResultsWe developed a weakly supervised deep learning model, achieving the AUCs of 0.748 (95% CI: 0.642–0.837), 0.729 (95% CI: 0.614–0.824) and 0.716 (95% CI: 0.617–0.802) on validation set and two external test sets, respectively. The DL-RRS provided a hazard ratio for high versus low recurrence risk of 2.67 (95% CI 1.62–4.38; p=0.0026) in the survival analysis of the external test cohort (IDDF2022-ABS-0265 Figure 1. RFS of patients on validation set and external test set).ConclusionsWe established and validated an effective deep learning model, which automatically extracts the histological microscopic features related to prognosis, has a good predictive performance of postoperative recurrence in CRC, and will better serve the clinical decision-making.